Rare bioparticle detection via deep metric learning.

Area of Science:

• Biomedical engineering
• Environmental science
• Computer science

Background:

• Deep neural networks excel in bioimage analysis but struggle with imbalanced datasets and rare event detection.
• Conventional classifiers like SoftMax have limitations in achieving both low false alarm and high recovery rates for critical applications.
• Deep metric learning enhances model generalizability by mapping images to a latent space, crucial for identifying rare objects.

Purpose of the Study:

• To develop a robust deep metric neural network model for detecting rare bioparticles (Cryptosporidium and Giardia) in drinking water.
• To address the challenges of imbalanced training data and variations in input images during inference.
• To improve the reliability of biosensing applications through enhanced bioparticle detection.

Main Methods:

• Implementation of a deep metric neural network architecture.
• Training the model on bioimage data, focusing on rare bioparticle identification.
• Utilizing distance information in a latent space to model image similarity for robust classification.

Main Results:

• The deep metric neural network achieved 99.86% classification accuracy.
• Demonstrated a precision rate of 98.89% and a recall rate of 99.16%.
• The model successfully achieved a zero false alarm rate in detecting rare bioparticles.

Conclusions:

• The proposed deep metric neural network offers a robust solution for rare bioparticle detection in environmental monitoring.
• The model's high performance empowers imaging flow cytometry for applications in biomedical diagnosis and biosensing.
• This approach significantly enhances the capability to detect contaminants like Cryptosporidium and Giardia in drinking water.